NASA and IBM have launched an open-source AI foundation model designed to process 17 years of lunar orbiter data for scientific research.

NASA and IBM Research have released the NASA-IBM Lunar Foundation Model, an open-source tool designed to make decades of lunar observation data more accessible for machine learning. The model is specifically optimized for tasks such as predicting polar ice deposits and detecting craters. It was developed through a collaboration between the two organizations and several academic institutions to address the challenge of utilizing vast amounts of unlabeled lunar data.

Training and Data Composition

The model was trained from scratch using SomBench, a multimodal lunar corpus containing nearly 2 million tile bundles. This dataset incorporates 17 years of observations from the Lunar Reconnaissance Orbiter (LRO), alongside data from the GRAIL, Lunar Prospector, and JAXA's Kaguya/SELENE missions. In total, the collection includes over 30 spatially aligned data layers derived from nine instruments.

Performance and Capabilities

In testing, the model demonstrated significant improvements over existing baselines, particularly in predicting polar ice deposits, where it reduced prediction error by up to 22 percent compared to the SwinV2-B baseline. For coarse-scale crater detection, the model outperformed the same baseline by nearly 19 percent using only half the training data. While the model excels at identifying patterns, researchers noted it is not intended for absolute geodetic positioning, as it can produce errors in latitude, longitude, and absolute height values.

Availability and Collaboration

The project is part of a broader 'AI for Science' collaboration between NASA and IBM, which has been active under a Space Act Agreement since early 2022. The lunar model is now publicly available on Hugging Face, with code hosted on GitHub and integration provided through the TerraTorch toolkit. The team also released the associated pretraining datasets and benchmark collections.